This pages lists the open BSc. and MSc. thesis descriptions, as well as the master projects opportunities currently available in the DDIS research group. Do not hesitate to contact the respective person if you are interested in one of the topics. If you would like to write a thesis about your own idea you can propose it to the person most related to what you plan to do or you can contact email@example.com directly.
Large-scale Active Learning for Concept Detection in Video
Computer Vision has made significant advances in recent years, resulting in, among other things, methods to detect and even localize objects in images and videos with high precision. These modern methods generally rely on large neural network architectures which are trained using large amounts of labelled training data. Such methods work well for situations where the visual data they operate on can be constrained to the same setting as was present in the training data. Their performance does however start to degrade if they are presented with input which is substantially different from what has been contained in the training set. Another limitation is that the addition of further classes to be detected involves re-training the network and can negatively influence the recognition performance for existing concepts. It is therefore rather uncommon to continuously adapt the weights of a specific network. In contrast, active learning algorithms incorporate a feedback loop in order to be able to continuously gather new information from external sources, commonly human supervisors.
The aim of this thesis is to design and implement a pipeline which uses a combination of state of the art deep neural network approaches in combination with active learning strategies in order to detect and precisely annotate an arbitrary number of objects in unconstrained images and videos. The pipeline should use neural network-based methods for the localization of objects and the embedding of their visual appearance into an appropriate vector representation which enables a subsequent bank of independent classifiers to identify the object. An active learning loop should be used to collect labels for objects which have not been seen before or can not be accurately classified in order to continuously update and extend the classifier bank and extend the number of classifiable objects.
This project is suitable for a Master Thesis and requires some prior familiarity with classical machine learning methods and deep learning as well as strong programming skills in Java. Python programming skills are strongly recommended.
Contact: Luca Rossetto